lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagevec.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static inline bool osd_use_page_cache(struct osd_device *d)
  60 {
  61         /* do not use pagecache if write and read caching are disabled */
  62         if (d->od_writethrough_cache + d->od_read_cache == 0)
  63                 return false;
  64         /* use pagecache by default */
  65         return true;
  66 }
  67
  68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  69                             int rw, int line, int pages)
  70 {
  71         int blocks, i;
  72
  73         LASSERTF(iobuf->dr_elapsed_valid == 0,
  74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  76                  iobuf->dr_init_at);
  77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  78
  79         init_waitqueue_head(&iobuf->dr_wait);
  80         atomic_set(&iobuf->dr_numreqs, 0);
  81         iobuf->dr_npages = 0;
  82         iobuf->dr_error = 0;
  83         iobuf->dr_dev = d;
  84         iobuf->dr_frags = 0;
  85         iobuf->dr_elapsed = ktime_set(0, 0);
  86         /* must be counted before, so assert */
  87         iobuf->dr_rw = rw;
  88         iobuf->dr_init_at = line;
  89
  90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  93                         pages * sizeof(iobuf->dr_pages[0]));
  94                 return 0;
  95         }
  96
  97         /* start with 1MB for 4K blocks */
  98         i = 256;
  99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 100                 i <<= 1;
 101
 102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 103                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 104         pages = i;
 105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 106         iobuf->dr_max_pages = 0;
 107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 108                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 109
 110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 112         if (unlikely(iobuf->dr_blocks == NULL))
 113                 return -ENOMEM;
 114
 115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 117         if (unlikely(iobuf->dr_pages == NULL))
 118                 return -ENOMEM;
 119
 120         lu_buf_realloc(&iobuf->dr_lnb_buf,
 121                        pages * sizeof(iobuf->dr_lnbs[0]));
 122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 123         if (unlikely(iobuf->dr_lnbs == NULL))
 124                 return -ENOMEM;
 125
 126         iobuf->dr_max_pages = pages;
 127
 128         return 0;
 129 }
 130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 132
 133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 134                                struct niobuf_local *lnb)
 135 {
 136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 139         iobuf->dr_npages++;
 140 }
 141
 142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 143 {
 144         int rw = iobuf->dr_rw;
 145
 146         if (iobuf->dr_elapsed_valid) {
 147                 iobuf->dr_elapsed_valid = 0;
 148                 LASSERT(iobuf->dr_dev == d);
 149                 LASSERT(iobuf->dr_frags > 0);
 150                 lprocfs_oh_tally(&d->od_brw_stats.
 151                                  hist[BRW_R_DIO_FRAGS+rw],
 152                                  iobuf->dr_frags);
 153                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 154                                       ktime_to_ms(iobuf->dr_elapsed));
 155         }
 156 }
 157
 158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 159 static void dio_complete_routine(struct bio *bio)
 160 {
 161         int error = bio->bi_status;
 162 #else
 163 static void dio_complete_routine(struct bio *bio, int error)
 164 {
 165 #endif
 166         struct osd_iobuf *iobuf = bio->bi_private;
 167         struct bio_vec *bvl;
 168
 169         /* CAVEAT EMPTOR: possibly in IRQ context
 170          * DO NOT record procfs stats here!!! */
 171
 172         if (unlikely(iobuf == NULL)) {
 173                 CERROR("***** bio->bi_private is NULL!  This should never "
 174                        "happen.  Normally, I would crash here, but instead I "
 175                        "will dump the bio contents to the console.  Please "
 176                        "report this to <https://jira.whamcloud.com/> , along "
 177                        "with any interesting messages leading up to this point "
 178                        "(like SCSI errors, perhaps).  Because bi_private is "
 179                        "NULL, I can't wake up the thread that initiated this "
 180                        "IO - you will probably have to reboot this node.\n");
 181                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 182                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 183                        bio->bi_next, (unsigned long)bio->bi_flags,
 184                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 185                        bio_sectors(bio) << 9, bio->bi_end_io,
 186                        atomic_read(&bio->__bi_cnt),
 187                        bio->bi_private);
 188                 return;
 189         }
 190
 191         /* the check is outside of the cycle for performance reason -bzzz */
 192         if (!bio_data_dir(bio)) {
 193                 DECLARE_BVEC_ITER_ALL(iter_all);
 194
 195                 bio_for_each_segment_all(bvl, bio, iter_all) {
 196                         if (likely(error == 0))
 197                                 SetPageUptodate(bvl_to_page(bvl));
 198                         LASSERT(PageLocked(bvl_to_page(bvl)));
 199                 }
 200                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 201         } else {
 202                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 203         }
 204
 205         /* any real error is good enough -bzzz */
 206         if (error != 0 && iobuf->dr_error == 0)
 207                 iobuf->dr_error = error;
 208
 209         /*
 210          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 211          * it's possible that service thread will see dr_numreqs
 212          * is zero, but dr_elapsed is not set yet, leading to lost
 213          * data in this processing and an assertion in a subsequent
 214          * call to OSD.
 215          */
 216         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 217                 ktime_t now = ktime_get();
 218
 219                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 220                 iobuf->dr_elapsed_valid = 1;
 221         }
 222         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 223                 wake_up(&iobuf->dr_wait);
 224
 225         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 226          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 227          * mempool when serious on-disk fragmentation was encountered,
 228          * deadlocking the OST.  The bios are now released as soon as complete
 229          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 230         bio_put(bio);
 231 }
 232
 233 static void record_start_io(struct osd_iobuf *iobuf, int size)
 234 {
 235         struct osd_device    *osd = iobuf->dr_dev;
 236         struct obd_histogram *h = osd->od_brw_stats.hist;
 237
 238         iobuf->dr_frags++;
 239         atomic_inc(&iobuf->dr_numreqs);
 240
 241         if (iobuf->dr_rw == 0) {
 242                 atomic_inc(&osd->od_r_in_flight);
 243                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 244                                  atomic_read(&osd->od_r_in_flight));
 245                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 246         } else if (iobuf->dr_rw == 1) {
 247                 atomic_inc(&osd->od_w_in_flight);
 248                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 249                                  atomic_read(&osd->od_w_in_flight));
 250                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 251         } else {
 252                 LBUG();
 253         }
 254 }
 255
 256 static void osd_submit_bio(int rw, struct bio *bio)
 257 {
 258         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 259 #ifdef HAVE_SUBMIT_BIO_2ARGS
 260         submit_bio(rw ? WRITE : READ, bio);
 261 #else
 262         bio->bi_opf |= rw;
 263         submit_bio(bio);
 264 #endif
 265 }
 266
 267 static int can_be_merged(struct bio *bio, sector_t sector)
 268 {
 269         if (bio == NULL)
 270                 return 0;
 271
 272         return bio_end_sector(bio) == sector ? 1 : 0;
 273 }
 274
 275 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 276 /*
 277  * This function will change the data written, thus it should only be
 278  * used when checking data integrity feature
 279  */
 280 static void bio_integrity_fault_inject(struct bio *bio)
 281 {
 282         struct bio_vec *bvec;
 283         DECLARE_BVEC_ITER_ALL(iter_all);
 284         void *kaddr;
 285         char *addr;
 286
 287         bio_for_each_segment_all(bvec, bio, iter_all) {
 288                 struct page *page = bvec->bv_page;
 289
 290                 kaddr = kmap(page);
 291                 addr = kaddr;
 292                 *addr = ~(*addr);
 293                 kunmap(page);
 294                 break;
 295         }
 296 }
 297
 298 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 299                            unsigned int sectors, int tuple_size)
 300 {
 301         __u16 *expected_guard;
 302         __u16 *bio_guard;
 303         int i;
 304
 305         expected_guard = expected_guard_buf;
 306         for (i = 0; i < sectors; i++) {
 307                 bio_guard = (__u16 *)bio_prot_buf;
 308                 if (*bio_guard != *expected_guard) {
 309                         CERROR("unexpected guard tags on sector %d "
 310                                "expected guard %u, bio guard "
 311                                "%u, sectors %u, tuple size %d\n",
 312                                i, *expected_guard, *bio_guard, sectors,
 313                                tuple_size);
 314                         return -EIO;
 315                 }
 316                 expected_guard++;
 317                 bio_prot_buf += tuple_size;
 318         }
 319         return 0;
 320 }
 321
 322 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 323                                      struct osd_iobuf *iobuf, int index)
 324 {
 325         struct blk_integrity *bi = bdev_get_integrity(bdev);
 326         struct bio_integrity_payload *bip = bio->bi_integrity;
 327         struct niobuf_local *lnb;
 328         unsigned short sector_size = blk_integrity_interval(bi);
 329         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 330                 bip->bip_vec->bv_offset;
 331         struct bio_vec *bv;
 332         sector_t sector = bio_start_sector(bio);
 333         unsigned int sectors, total;
 334         DECLARE_BVEC_ITER_ALL(iter_all);
 335         __u16 *expected_guard;
 336         int rc;
 337
 338         total = 0;
 339         bio_for_each_segment_all(bv, bio, iter_all) {
 340                 lnb = iobuf->dr_lnbs[index];
 341                 expected_guard = lnb->lnb_guards;
 342                 sectors = bv->bv_len / sector_size;
 343                 if (lnb->lnb_guard_rpc) {
 344                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 345                                              sectors, bi->tuple_size);
 346                         if (rc)
 347                                 return rc;
 348                 }
 349
 350                 sector += sectors;
 351                 bio_prot_buf += sectors * bi->tuple_size;
 352                 total += sectors * bi->tuple_size;
 353                 LASSERT(total <= bip_size(bio->bi_integrity));
 354                 index++;
 355         }
 356         return 0;
 357 }
 358
 359 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 360                                     struct osd_iobuf *iobuf,
 361                                     int start_page_idx, bool fault_inject,
 362                                     bool integrity_enabled)
 363 {
 364         struct super_block *sb = osd_sb(osd);
 365         integrity_gen_fn *generate_fn = NULL;
 366         integrity_vrfy_fn *verify_fn = NULL;
 367         int rc;
 368
 369         ENTRY;
 370
 371         if (!integrity_enabled)
 372                 RETURN(0);
 373
 374         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 375         if (rc)
 376                 RETURN(rc);
 377
 378         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 379         if (rc)
 380                 RETURN(rc);
 381
 382         /* Verify and inject fault only when writing */
 383         if (iobuf->dr_rw == 1) {
 384                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 385                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 386                                                        start_page_idx);
 387                         if (rc)
 388                                 RETURN(rc);
 389                 }
 390
 391                 if (unlikely(fault_inject))
 392                         bio_integrity_fault_inject(bio);
 393         }
 394
 395         RETURN(0);
 396 }
 397
 398 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 399 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 400 static void dio_integrity_complete_routine(struct bio *bio)
 401 #  else
 402 static void dio_integrity_complete_routine(struct bio *bio, int error)
 403 #  endif
 404 {
 405         struct osd_bio_private *bio_private = bio->bi_private;
 406
 407         bio->bi_private = bio_private->obp_iobuf;
 408         osd_dio_complete_routine(bio, error);
 409
 410         OBD_FREE_PTR(bio_private);
 411 }
 412 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 413 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 414 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 415                                  fault_inject, integrity_enabled) 0
 416 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 417
 418 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 419                         bool integrity_enabled, int start_page_idx,
 420                         struct osd_bio_private **pprivate)
 421 {
 422         ENTRY;
 423
 424         *pprivate = NULL;
 425
 426 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 427         if (integrity_enabled) {
 428                 struct osd_bio_private *bio_private = NULL;
 429
 430                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 431                 if (bio_private == NULL)
 432                         RETURN(-ENOMEM);
 433                 bio->bi_end_io = dio_integrity_complete_routine;
 434                 bio->bi_private = bio_private;
 435                 bio_private->obp_start_page_idx = start_page_idx;
 436                 bio_private->obp_iobuf = iobuf;
 437                 *pprivate = bio_private;
 438         } else
 439 #endif
 440         {
 441                 bio->bi_end_io = dio_complete_routine;
 442                 bio->bi_private = iobuf;
 443         }
 444
 445         RETURN(0);
 446 }
 447
 448 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 449                       struct osd_iobuf *iobuf)
 450 {
 451         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 452         struct page **pages = iobuf->dr_pages;
 453         int npages = iobuf->dr_npages;
 454         sector_t *blocks = iobuf->dr_blocks;
 455         int total_blocks = npages * blocks_per_page;
 456         struct super_block *sb = inode->i_sb;
 457         int sector_bits = sb->s_blocksize_bits - 9;
 458         unsigned int blocksize = sb->s_blocksize;
 459         struct block_device *bdev = sb->s_bdev;
 460         struct osd_bio_private *bio_private = NULL;
 461         struct bio *bio = NULL;
 462         int bio_start_page_idx;
 463         struct page *page;
 464         unsigned int page_offset;
 465         sector_t sector;
 466         int nblocks;
 467         int block_idx;
 468         int page_idx;
 469         int i;
 470         int rc = 0;
 471         bool fault_inject;
 472         bool integrity_enabled;
 473         struct blk_plug plug;
 474         ENTRY;
 475
 476         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 477         LASSERT(iobuf->dr_npages == npages);
 478
 479         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 480
 481         osd_brw_stats_update(osd, iobuf);
 482         iobuf->dr_start_time = ktime_get();
 483
 484         blk_start_plug(&plug);
 485         for (page_idx = 0, block_idx = 0;
 486              page_idx < npages;
 487              page_idx++, block_idx += blocks_per_page) {
 488
 489                 page = pages[page_idx];
 490                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 491
 492                 for (i = 0, page_offset = 0;
 493                      i < blocks_per_page;
 494                      i += nblocks, page_offset += blocksize * nblocks) {
 495
 496                         nblocks = 1;
 497
 498                         if (blocks[block_idx + i] == 0) {  /* hole */
 499                                 LASSERTF(iobuf->dr_rw == 0,
 500                                          "page_idx %u, block_idx %u, i %u\n",
 501                                          page_idx, block_idx, i);
 502                                 memset(kmap(page) + page_offset, 0, blocksize);
 503                                 kunmap(page);
 504                                 continue;
 505                         }
 506
 507                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 508
 509                         /* Additional contiguous file blocks? */
 510                         while (i + nblocks < blocks_per_page &&
 511                                (sector + (nblocks << sector_bits)) ==
 512                                ((sector_t)blocks[block_idx + i + nblocks] <<
 513                                 sector_bits))
 514                                 nblocks++;
 515
 516                         if (bio != NULL &&
 517                             can_be_merged(bio, sector) &&
 518                             bio_add_page(bio, page,
 519                                          blocksize * nblocks, page_offset) != 0)
 520                                 continue;       /* added this frag OK */
 521
 522                         if (bio != NULL) {
 523                                 struct request_queue *q = bio_get_queue(bio);
 524                                 unsigned int bi_size = bio_sectors(bio) << 9;
 525
 526                                 /* Dang! I have to fragment this I/O */
 527                                 CDEBUG(D_INODE,
 528                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 529                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 530                                        bio_sectors(bio),
 531                                        queue_max_sectors(q),
 532                                        osd_bio_nr_segs(bio),
 533                                        queue_max_segments(q));
 534                                 rc = osd_bio_integrity_handle(osd, bio,
 535                                         iobuf, bio_start_page_idx,
 536                                         fault_inject, integrity_enabled);
 537                                 if (rc) {
 538                                         bio_put(bio);
 539                                         goto out;
 540                                 }
 541
 542                                 record_start_io(iobuf, bi_size);
 543                                 osd_submit_bio(iobuf->dr_rw, bio);
 544                         }
 545
 546                         bio_start_page_idx = page_idx;
 547                         /* allocate new bio */
 548                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 549                                                       (npages - page_idx) *
 550                                                       blocks_per_page));
 551                         if (bio == NULL) {
 552                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 553                                        (npages - page_idx), blocks_per_page,
 554                                        (npages - page_idx) * blocks_per_page);
 555                                 rc = -ENOMEM;
 556                                 goto out;
 557                         }
 558
 559                         bio_set_dev(bio, bdev);
 560                         bio_set_sector(bio, sector);
 561                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 562                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 563                                           bio_start_page_idx, &bio_private);
 564                         if (rc) {
 565                                 bio_put(bio);
 566                                 goto out;
 567                         }
 568
 569                         rc = bio_add_page(bio, page,
 570                                           blocksize * nblocks, page_offset);
 571                         LASSERT(rc != 0);
 572                 }
 573         }
 574
 575         if (bio != NULL) {
 576                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 577                                               bio_start_page_idx,
 578                                               fault_inject,
 579                                               integrity_enabled);
 580                 if (rc) {
 581                         bio_put(bio);
 582                         goto out;
 583                 }
 584
 585                 record_start_io(iobuf, bio_sectors(bio) << 9);
 586                 osd_submit_bio(iobuf->dr_rw, bio);
 587                 rc = 0;
 588         }
 589
 590 out:
 591         blk_finish_plug(&plug);
 592
 593         /* in order to achieve better IO throughput, we don't wait for writes
 594          * completion here. instead we proceed with transaction commit in
 595          * parallel and wait for IO completion once transaction is stopped
 596          * see osd_trans_stop() for more details -bzzz */
 597         if (iobuf->dr_rw == 0 || fault_inject) {
 598                 wait_event(iobuf->dr_wait,
 599                            atomic_read(&iobuf->dr_numreqs) == 0);
 600                 osd_fini_iobuf(osd, iobuf);
 601         }
 602
 603         if (rc == 0) {
 604                 rc = iobuf->dr_error;
 605         } else {
 606                 if (bio_private)
 607                         OBD_FREE_PTR(bio_private);
 608         }
 609
 610         RETURN(rc);
 611 }
 612
 613 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 614                                    struct niobuf_local *lnb, int maxlnb)
 615 {
 616         int rc = 0;
 617         ENTRY;
 618
 619         *nrpages = 0;
 620
 621         while (len > 0) {
 622                 int poff = offset & (PAGE_SIZE - 1);
 623                 int plen = PAGE_SIZE - poff;
 624
 625                 if (*nrpages >= maxlnb) {
 626                         rc = -EOVERFLOW;
 627                         break;
 628                 }
 629
 630                 if (plen > len)
 631                         plen = len;
 632                 lnb->lnb_file_offset = offset;
 633                 lnb->lnb_page_offset = poff;
 634                 lnb->lnb_len = plen;
 635                 /* lnb->lnb_flags = rnb->rnb_flags; */
 636                 lnb->lnb_flags = 0;
 637                 lnb->lnb_page = NULL;
 638                 lnb->lnb_rc = 0;
 639                 lnb->lnb_guard_rpc = 0;
 640                 lnb->lnb_guard_disk = 0;
 641                 lnb->lnb_locked = 0;
 642
 643                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 644                          (long long) len);
 645                 offset += plen;
 646                 len -= plen;
 647                 lnb++;
 648                 (*nrpages)++;
 649         }
 650
 651         RETURN(rc);
 652 }
 653
 654 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 655                                  loff_t offset, gfp_t gfp_mask, bool cache)
 656 {
 657         struct osd_thread_info *oti = osd_oti_get(env);
 658         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 659         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 660         struct page *page;
 661         int cur;
 662
 663         LASSERT(inode);
 664
 665         if (cache) {
 666                 page = find_or_create_page(inode->i_mapping,
 667                                            offset >> PAGE_SHIFT, gfp_mask);
 668
 669                 if (likely(page)) {
 670                         LASSERT(!PagePrivate2(page));
 671                         wait_on_page_writeback(page);
 672                 } else {
 673                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 674                 }
 675
 676                 return page;
 677         }
 678
 679         if (inode->i_mapping->nrpages) {
 680                 /* consult with pagecache, but do not create new pages */
 681                 /* this is normally used once */
 682                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 683                 if (page) {
 684                         wait_on_page_writeback(page);
 685                         return page;
 686                 }
 687         }
 688
 689         LASSERT(oti->oti_dio_pages);
 690         cur = oti->oti_dio_pages_used;
 691         page = oti->oti_dio_pages[cur];
 692
 693         if (unlikely(!page)) {
 694                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 695                 page = alloc_page(gfp_mask);
 696                 if (!page)
 697                         return NULL;
 698                 oti->oti_dio_pages[cur] = page;
 699                 SetPagePrivate2(page);
 700                 lock_page(page);
 701         }
 702
 703         ClearPageUptodate(page);
 704         page->index = offset >> PAGE_SHIFT;
 705         oti->oti_dio_pages_used++;
 706
 707         return page;
 708 }
 709
 710 /*
 711  * there are following "locks":
 712  * journal_start
 713  * i_mutex
 714  * page lock
 715  *
 716  * osd write path:
 717  *  - lock page(s)
 718  *  - journal_start
 719  *  - truncate_sem
 720  *
 721  * ext4 vmtruncate:
 722  *  - lock pages, unlock
 723  *  - journal_start
 724  *  - lock partial page
 725  *  - i_data_sem
 726  *
 727  */
 728
 729 /**
 730  * Unlock and release pages loaded by osd_bufs_get()
 731  *
 732  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 733  *
 734  * \param env           thread execution environment
 735  * \param dt            dt object undergoing IO (OSD object + methods)
 736  * \param lnb           array of pages undergoing IO
 737  * \param npages        number of pages in \a lnb
 738  *
 739  * \retval 0            always
 740  */
 741 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 742                         struct niobuf_local *lnb, int npages)
 743 {
 744         struct osd_thread_info *oti = osd_oti_get(env);
 745         struct pagevec pvec;
 746         int i;
 747
 748         ll_pagevec_init(&pvec, 0);
 749
 750         for (i = 0; i < npages; i++) {
 751                 struct page *page = lnb[i].lnb_page;
 752
 753                 if (page == NULL)
 754                         continue;
 755
 756                 /* if the page isn't cached, then reset uptodate
 757                  * to prevent reuse */
 758                 if (PagePrivate2(page)) {
 759                         oti->oti_dio_pages_used--;
 760                 } else {
 761                         if (lnb[i].lnb_locked)
 762                                 unlock_page(page);
 763                         if (pagevec_add(&pvec, page) == 0)
 764                                 pagevec_release(&pvec);
 765                 }
 766
 767                 lnb[i].lnb_page = NULL;
 768         }
 769
 770         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 771
 772         /* Release any partial pagevec */
 773         pagevec_release(&pvec);
 774
 775         RETURN(0);
 776 }
 777
 778 /**
 779  * Load and lock pages undergoing IO
 780  *
 781  * Pages as described in the \a lnb array are fetched (from disk or cache)
 782  * and locked for IO by the caller.
 783  *
 784  * DLM locking protects us from write and truncate competing for same region,
 785  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 786  * It's possible the writeout on a such a page is in progress when we access
 787  * it. It's also possible that during this writeout we put new (partial) data
 788  * into the page, but won't be able to proceed in filter_commitrw_write().
 789  * Therefore, just wait for writeout completion as it should be rare enough.
 790  *
 791  * \param env           thread execution environment
 792  * \param dt            dt object undergoing IO (OSD object + methods)
 793  * \param pos           byte offset of IO start
 794  * \param len           number of bytes of IO
 795  * \param lnb           array of extents undergoing IO
 796  * \param rw            read or write operation, and other flags
 797  * \param capa          capabilities
 798  *
 799  * \retval pages        (zero or more) loaded successfully
 800  * \retval -ENOMEM      on memory/page allocation error
 801  */
 802 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 803                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 804                         int maxlnb, enum dt_bufs_type rw)
 805 {
 806         struct osd_thread_info *oti = osd_oti_get(env);
 807         struct osd_object *obj = osd_dt_obj(dt);
 808         struct osd_device *osd   = osd_obj2dev(obj);
 809         int npages, i, iosize, rc = 0;
 810         bool cache, write;
 811         loff_t fsize;
 812         gfp_t gfp_mask;
 813
 814         LASSERT(obj->oo_inode);
 815
 816         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 817         if (rc)
 818                 RETURN(rc);
 819
 820         write = rw & DT_BUFS_TYPE_WRITE;
 821
 822         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 823         iosize = fsize - lnb[0].lnb_file_offset;
 824         fsize = max(fsize, i_size_read(obj->oo_inode));
 825
 826         cache = rw & DT_BUFS_TYPE_READAHEAD;
 827         if (cache)
 828                 goto bypass_checks;
 829
 830         cache = osd_use_page_cache(osd);
 831         while (cache) {
 832                 if (write) {
 833                         if (!osd->od_writethrough_cache) {
 834                                 cache = false;
 835                                 break;
 836                         }
 837                         if (iosize > osd->od_writethrough_max_iosize) {
 838                                 cache = false;
 839                                 break;
 840                         }
 841                 } else {
 842                         if (!osd->od_read_cache) {
 843                                 cache = false;
 844                                 break;
 845                         }
 846                         if (iosize > osd->od_readcache_max_iosize) {
 847                                 cache = false;
 848                                 break;
 849                         }
 850                 }
 851                 /* don't use cache on large files */
 852                 if (osd->od_readcache_max_filesize &&
 853                     fsize > osd->od_readcache_max_filesize)
 854                         cache = false;
 855                 break;
 856         }
 857
 858 bypass_checks:
 859         if (!cache && unlikely(!oti->oti_dio_pages)) {
 860                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 861                                           PTLRPC_MAX_BRW_PAGES);
 862                 if (!oti->oti_dio_pages)
 863                         return -ENOMEM;
 864         }
 865
 866         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 867         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 868                                              GFP_HIGHUSER;
 869         for (i = 0; i < npages; i++, lnb++) {
 870                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 871                                              gfp_mask, cache);
 872                 if (lnb->lnb_page == NULL)
 873                         GOTO(cleanup, rc = -ENOMEM);
 874
 875                 lnb->lnb_locked = 1;
 876         }
 877
 878 #if 0
 879         /* XXX: this version doesn't invalidate cached pages, but use them */
 880         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 881                 /* do not allow data aliasing, invalidate pagecache */
 882                 /* XXX: can be quite expensive in mixed case */
 883                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 884                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 885                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 886         }
 887 #endif
 888
 889         RETURN(i);
 890
 891 cleanup:
 892         if (i > 0)
 893                 osd_bufs_put(env, dt, lnb - i, i);
 894         return rc;
 895 }
 896
 897 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 898                                        int pages, sector_t *blocks,
 899                                        int create)
 900 {
 901         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 902         int rc = 0, i = 0;
 903         struct page *fp = NULL;
 904         int clen = 0;
 905         pgoff_t max_page_index;
 906         handle_t *handle = NULL;
 907
 908         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 909
 910         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 911                 inode->i_ino, pages, (*page)->index);
 912
 913         if (create) {
 914                 create = LDISKFS_GET_BLOCKS_CREATE;
 915                 handle = ldiskfs_journal_current_handle();
 916                 LASSERT(handle != NULL);
 917                 rc = osd_attach_jinode(inode);
 918                 if (rc)
 919                         return rc;
 920         }
 921         /* pages are sorted already. so, we just have to find
 922          * contig. space and process them properly */
 923         while (i < pages) {
 924                 long blen, total = 0;
 925                 struct ldiskfs_map_blocks map = { 0 };
 926
 927                 if (fp == NULL) { /* start new extent */
 928                         fp = *page++;
 929                         clen = 1;
 930                         if (++i != pages)
 931                                 continue;
 932                 } else if (fp->index + clen == (*page)->index) {
 933                         /* continue the extent */
 934                         page++;
 935                         clen++;
 936                         if (++i != pages)
 937                                 continue;
 938                 }
 939                 if (fp->index + clen >= max_page_index)
 940                         GOTO(cleanup, rc = -EFBIG);
 941                 /* process found extent */
 942                 map.m_lblk = fp->index * blocks_per_page;
 943                 map.m_len = blen = clen * blocks_per_page;
 944 cont_map:
 945                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 946                 if (rc >= 0) {
 947                         int c = 0;
 948                         for (; total < blen && c < map.m_len; c++, total++) {
 949                                 if (rc == 0) {
 950                                         *(blocks + total) = 0;
 951                                         total++;
 952                                         break;
 953                                 } else {
 954                                         *(blocks + total) = map.m_pblk + c;
 955                                         /* unmap any possible underlying
 956                                          * metadata from the block device
 957                                          * mapping.  bug 6998. */
 958                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 959                                             create)
 960                                                 clean_bdev_aliases(
 961                                                         inode->i_sb->s_bdev,
 962                                                         map.m_pblk + c, 1);
 963                                 }
 964                         }
 965                         rc = 0;
 966                 }
 967                 if (rc == 0 && total < blen) {
 968                         map.m_lblk = fp->index * blocks_per_page + total;
 969                         map.m_len = blen - total;
 970                         goto cont_map;
 971                 }
 972                 if (rc != 0)
 973                         GOTO(cleanup, rc);
 974
 975                 /* look for next extent */
 976                 fp = NULL;
 977                 blocks += blocks_per_page * clen;
 978         }
 979 cleanup:
 980         return rc;
 981 }
 982
 983 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 984                           struct niobuf_local *lnb, int npages)
 985 {
 986         struct osd_thread_info *oti   = osd_oti_get(env);
 987         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 988         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 989         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 990         ktime_t start, end;
 991         s64 timediff;
 992         ssize_t isize;
 993         __s64  maxidx;
 994         int i, rc = 0;
 995
 996         LASSERT(inode);
 997
 998         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 999         if (unlikely(rc != 0))
1000                 RETURN(rc);
1001
1002         isize = i_size_read(inode);
1003         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1004
1005         start = ktime_get();
1006         for (i = 0; i < npages; i++) {
1007
1008                 /*
1009                  * till commit the content of the page is undefined
1010                  * we'll set it uptodate once bulk is done. otherwise
1011                  * subsequent reads can access non-stable data
1012                  */
1013                 ClearPageUptodate(lnb[i].lnb_page);
1014
1015                 if (lnb[i].lnb_len == PAGE_SIZE)
1016                         continue;
1017
1018                 if (maxidx >= lnb[i].lnb_page->index) {
1019                         osd_iobuf_add_page(iobuf, &lnb[i]);
1020                 } else {
1021                         long off;
1022                         char *p = kmap(lnb[i].lnb_page);
1023
1024                         off = lnb[i].lnb_page_offset;
1025                         if (off)
1026                                 memset(p, 0, off);
1027                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1028                               ~PAGE_MASK;
1029                         if (off)
1030                                 memset(p + off, 0, PAGE_SIZE - off);
1031                         kunmap(lnb[i].lnb_page);
1032                 }
1033         }
1034         end = ktime_get();
1035         timediff = ktime_us_delta(end, start);
1036         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1037
1038         if (iobuf->dr_npages) {
1039                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1040                                                  iobuf->dr_npages,
1041                                                  iobuf->dr_blocks, 0);
1042                 if (likely(rc == 0)) {
1043                         rc = osd_do_bio(osd, inode, iobuf);
1044                         /* do IO stats for preparation reads */
1045                         osd_fini_iobuf(osd, iobuf);
1046                 }
1047         }
1048         RETURN(rc);
1049 }
1050
1051 struct osd_fextent {
1052         sector_t        start;
1053         sector_t        end;
1054         unsigned int    mapped:1;
1055 };
1056
1057 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1058                          struct osd_fextent *cached_extent)
1059 {
1060         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1061         sector_t block = offset >> inode->i_blkbits;
1062         sector_t start;
1063         struct fiemap_extent_info fei = { 0 };
1064         struct fiemap_extent fe = { 0 };
1065         mm_segment_t saved_fs;
1066         int rc;
1067
1068         if (block >= cached_extent->start && block < cached_extent->end)
1069                 return cached_extent->mapped;
1070
1071         if (i_size_read(inode) == 0)
1072                 return 0;
1073
1074         /* Beyond EOF, must not be mapped */
1075         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1076                 return 0;
1077
1078         fei.fi_extents_max = 1;
1079         fei.fi_extents_start = &fe;
1080
1081         saved_fs = get_fs();
1082         set_fs(KERNEL_DS);
1083         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1084         set_fs(saved_fs);
1085         if (rc != 0)
1086                 return 0;
1087
1088         start = fe.fe_logical >> inode->i_blkbits;
1089
1090         if (start > block) {
1091                 cached_extent->start = block;
1092                 cached_extent->end = start;
1093                 cached_extent->mapped = 0;
1094         } else {
1095                 cached_extent->start = start;
1096                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1097                                       inode->i_blkbits;
1098                 cached_extent->mapped = 1;
1099         }
1100
1101         return cached_extent->mapped;
1102 }
1103
1104 static int osd_declare_write_commit(const struct lu_env *env,
1105                                     struct dt_object *dt,
1106                                     struct niobuf_local *lnb, int npages,
1107                                     struct thandle *handle)
1108 {
1109         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1110         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1111         struct osd_thandle      *oh;
1112         int                     extents = 1;
1113         int                     depth;
1114         int                     i;
1115         int                     newblocks;
1116         int                     rc = 0;
1117         int                     credits = 0;
1118         long long               quota_space = 0;
1119         struct osd_fextent      extent = { 0 };
1120         enum osd_quota_local_flags local_flags = 0;
1121         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1122         ENTRY;
1123
1124         LASSERT(handle != NULL);
1125         oh = container_of(handle, struct osd_thandle, ot_super);
1126         LASSERT(oh->ot_handle == NULL);
1127
1128         newblocks = npages;
1129
1130         /* calculate number of extents (probably better to pass nb) */
1131         for (i = 0; i < npages; i++) {
1132                 if (i && lnb[i].lnb_file_offset !=
1133                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1134                         extents++;
1135
1136                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1137                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1138                 else
1139                         quota_space += PAGE_SIZE;
1140
1141                 /* ignore quota for the whole request if any page is from
1142                  * client cache or written by root.
1143                  *
1144                  * XXX once we drop the 1.8 client support, the checking
1145                  * for whether page is from cache can be simplified as:
1146                  * !(lnb[i].flags & OBD_BRW_SYNC)
1147                  *
1148                  * XXX we could handle this on per-lnb basis as done by
1149                  * grant. */
1150                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1151                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1152                     OBD_BRW_FROM_GRANT)
1153                         declare_flags |= OSD_QID_FORCE;
1154         }
1155
1156         /*
1157          * each extent can go into new leaf causing a split
1158          * 5 is max tree depth: inode + 4 index blocks
1159          * with blockmaps, depth is 3 at most
1160          */
1161         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1162                 /*
1163                  * many concurrent threads may grow tree by the time
1164                  * our transaction starts. so, consider 2 is a min depth
1165                  */
1166                 depth = ext_depth(inode);
1167                 depth = max(depth, 1) + 1;
1168                 newblocks += depth;
1169                 credits++; /* inode */
1170                 credits += depth * 2 * extents;
1171         } else {
1172                 depth = 3;
1173                 newblocks += depth;
1174                 credits++; /* inode */
1175                 credits += depth * extents;
1176         }
1177
1178         /* quota space for metadata blocks */
1179         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1180
1181         /* quota space should be reported in 1K blocks */
1182         quota_space = toqb(quota_space);
1183
1184         /* each new block can go in different group (bitmap + gd) */
1185
1186         /* we can't dirty more bitmap blocks than exist */
1187         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1188                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1189         else
1190                 credits += newblocks;
1191
1192         /* we can't dirty more gd blocks than exist */
1193         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1194                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1195         else
1196                 credits += newblocks;
1197
1198         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1199
1200         /* make sure the over quota flags were not set */
1201         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1202
1203         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1204                                    i_projid_read(inode), quota_space, oh,
1205                                    osd_dt_obj(dt), &local_flags, declare_flags);
1206
1207         /* we need only to store the overquota flags in the first lnb for
1208          * now, once we support multiple objects BRW, this code needs be
1209          * revised. */
1210         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1211                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1212         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1213                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1214         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1215                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1216
1217         if (rc == 0)
1218                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1219
1220         RETURN(rc);
1221 }
1222
1223 /* Check if a block is allocated or not */
1224 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1225                             struct niobuf_local *lnb, int npages,
1226                             struct thandle *thandle, __u64 user_size)
1227 {
1228         struct osd_thread_info *oti = osd_oti_get(env);
1229         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1230         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1231         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1232         loff_t disk_size;
1233         int rc = 0, i;
1234
1235         LASSERT(inode);
1236
1237         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1238         if (unlikely(rc != 0))
1239                 RETURN(rc);
1240
1241         disk_size = i_size_read(inode);
1242         /* if disk_size is already bigger than specified user_size,
1243          * ignore user_size
1244          */
1245         if (disk_size > user_size)
1246                 user_size = 0;
1247         dquot_initialize(inode);
1248
1249         for (i = 0; i < npages; i++) {
1250                 if (lnb[i].lnb_rc == -ENOSPC &&
1251                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1252                         /* Allow the write to proceed if overwriting an
1253                          * existing block */
1254                         lnb[i].lnb_rc = 0;
1255                 }
1256
1257                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1258                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1259                                lnb[i].lnb_rc);
1260                         LASSERT(lnb[i].lnb_page);
1261                         generic_error_remove_page(inode->i_mapping,
1262                                                   lnb[i].lnb_page);
1263                         continue;
1264                 }
1265
1266                 LASSERT(PageLocked(lnb[i].lnb_page));
1267                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1268
1269                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size)
1270                         disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1271
1272                 /*
1273                  * Since write and truncate are serialized by oo_sem, even
1274                  * partial-page truncate should not leave dirty pages in the
1275                  * page cache.
1276                  */
1277                 LASSERT(!PageDirty(lnb[i].lnb_page));
1278
1279                 SetPageUptodate(lnb[i].lnb_page);
1280
1281                 osd_iobuf_add_page(iobuf, &lnb[i]);
1282         }
1283         /* if file has grown, take user_size into account */
1284         if (user_size && disk_size > user_size)
1285                 disk_size = user_size;
1286
1287         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1288
1289         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1290                 rc = -ENOSPC;
1291         } else if (iobuf->dr_npages > 0) {
1292                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1293                                                  iobuf->dr_npages,
1294                                                  iobuf->dr_blocks, 1);
1295         } else {
1296                 /* no pages to write, no transno is needed */
1297                 thandle->th_local = 1;
1298         }
1299
1300         if (likely(rc == 0)) {
1301                 spin_lock(&inode->i_lock);
1302                 if (disk_size > i_size_read(inode)) {
1303                         i_size_write(inode, disk_size);
1304                         LDISKFS_I(inode)->i_disksize = disk_size;
1305                         spin_unlock(&inode->i_lock);
1306                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1307                 } else {
1308                         spin_unlock(&inode->i_lock);
1309                 }
1310
1311                 rc = osd_do_bio(osd, inode, iobuf);
1312                 /* we don't do stats here as in read path because
1313                  * write is async: we'll do this in osd_put_bufs() */
1314         } else {
1315                 osd_fini_iobuf(osd, iobuf);
1316         }
1317
1318         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1319
1320         if (unlikely(rc != 0)) {
1321                 /* if write fails, we should drop pages from the cache */
1322                 for (i = 0; i < npages; i++) {
1323                         if (lnb[i].lnb_page == NULL)
1324                                 continue;
1325                         if (!PagePrivate2(lnb[i].lnb_page)) {
1326                                 LASSERT(PageLocked(lnb[i].lnb_page));
1327                                 generic_error_remove_page(inode->i_mapping,
1328                                                           lnb[i].lnb_page);
1329                         }
1330                 }
1331         }
1332
1333         RETURN(rc);
1334 }
1335
1336 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1337                          struct niobuf_local *lnb, int npages)
1338 {
1339         struct osd_thread_info *oti = osd_oti_get(env);
1340         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1341         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1342         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1343         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1344         ktime_t start, end;
1345         s64 timediff;
1346         loff_t isize;
1347
1348         LASSERT(inode);
1349
1350         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1351         if (unlikely(rc != 0))
1352                 RETURN(rc);
1353
1354         isize = i_size_read(inode);
1355
1356         start = ktime_get();
1357         for (i = 0; i < npages; i++) {
1358
1359                 if (isize <= lnb[i].lnb_file_offset)
1360                         /* If there's no more data, abort early.
1361                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1362                         break;
1363
1364                 /* instead of looking if we go beyong isize, send complete
1365                  * pages all the time
1366                  */
1367                 lnb[i].lnb_rc = lnb[i].lnb_len;
1368
1369                 /* Bypass disk read if fail_loc is set properly */
1370                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1371                         SetPageUptodate(lnb[i].lnb_page);
1372
1373                 if (PageUptodate(lnb[i].lnb_page)) {
1374                         cache_hits++;
1375                         unlock_page(lnb[i].lnb_page);
1376                 } else {
1377                         cache_misses++;
1378                         osd_iobuf_add_page(iobuf, &lnb[i]);
1379                 }
1380                 /* no need to unlock in osd_bufs_put(), the sooner page is
1381                  * unlocked, the earlier another client can access it.
1382                  * notice real unlock_page() can be called few lines
1383                  * below after osd_do_bio(). lnb is a per-thread, so it's
1384                  * fine to have PG_locked and lnb_locked inconsistent here */
1385                 lnb[i].lnb_locked = 0;
1386         }
1387         end = ktime_get();
1388         timediff = ktime_us_delta(end, start);
1389         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1390
1391         if (cache_hits != 0)
1392                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1393                                     cache_hits);
1394         if (cache_misses != 0)
1395                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1396                                     cache_misses);
1397         if (cache_hits + cache_misses != 0)
1398                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1399                                     cache_hits + cache_misses);
1400
1401         if (iobuf->dr_npages) {
1402                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1403                                                  iobuf->dr_npages,
1404                                                  iobuf->dr_blocks, 0);
1405                 rc = osd_do_bio(osd, inode, iobuf);
1406
1407                 /* IO stats will be done in osd_bufs_put() */
1408
1409                 /* early release to let others read data during the bulk */
1410                 for (i = 0; i < iobuf->dr_npages; i++) {
1411                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1412                         if (!PagePrivate2(iobuf->dr_pages[i]))
1413                                 unlock_page(iobuf->dr_pages[i]);
1414                 }
1415         }
1416
1417         RETURN(rc);
1418 }
1419
1420 /*
1421  * XXX: Another layering violation for now.
1422  *
1423  * We don't want to use ->f_op->read methods, because generic file write
1424  *
1425  *         - serializes on ->i_sem, and
1426  *
1427  *         - does a lot of extra work like balance_dirty_pages(),
1428  *
1429  * which doesn't work for globally shared files like /last_rcvd.
1430  */
1431 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1432 {
1433         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1434
1435         memcpy(buffer, (char *)ei->i_data, buflen);
1436
1437         return  buflen;
1438 }
1439
1440 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1441 {
1442         struct buffer_head *bh;
1443         unsigned long block;
1444         int osize;
1445         int blocksize;
1446         int csize;
1447         int boffs;
1448
1449         /* prevent reading after eof */
1450         spin_lock(&inode->i_lock);
1451         if (i_size_read(inode) < *offs + size) {
1452                 loff_t diff = i_size_read(inode) - *offs;
1453                 spin_unlock(&inode->i_lock);
1454                 if (diff < 0) {
1455                         CDEBUG(D_OTHER,
1456                                "size %llu is too short to read @%llu\n",
1457                                i_size_read(inode), *offs);
1458                         return -EBADR;
1459                 } else if (diff == 0) {
1460                         return 0;
1461                 } else {
1462                         size = diff;
1463                 }
1464         } else {
1465                 spin_unlock(&inode->i_lock);
1466         }
1467
1468         blocksize = 1 << inode->i_blkbits;
1469         osize = size;
1470         while (size > 0) {
1471                 block = *offs >> inode->i_blkbits;
1472                 boffs = *offs & (blocksize - 1);
1473                 csize = min(blocksize - boffs, size);
1474                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1475                 if (IS_ERR(bh)) {
1476                         CERROR("%s: can't read %u@%llu on ino %lu: "
1477                                "rc = %ld\n", osd_ino2name(inode),
1478                                csize, *offs, inode->i_ino,
1479                                PTR_ERR(bh));
1480                         return PTR_ERR(bh);
1481                 }
1482
1483                 if (bh != NULL) {
1484                         memcpy(buf, bh->b_data + boffs, csize);
1485                         brelse(bh);
1486                 } else {
1487                         memset(buf, 0, csize);
1488                 }
1489
1490                 *offs += csize;
1491                 buf += csize;
1492                 size -= csize;
1493         }
1494         return osize;
1495 }
1496
1497 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1498                         struct lu_buf *buf, loff_t *pos)
1499 {
1500         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1501         int rc;
1502
1503         /* Read small symlink from inode body as we need to maintain correct
1504          * on-disk symlinks for ldiskfs.
1505          */
1506         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1507                 loff_t size = i_size_read(inode);
1508
1509                 if (buf->lb_len < size)
1510                         return -EOVERFLOW;
1511
1512                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1513                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1514                 else
1515                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1516         } else {
1517                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1518         }
1519
1520         return rc;
1521 }
1522
1523 static inline int osd_extents_enabled(struct super_block *sb,
1524                                       struct inode *inode)
1525 {
1526         if (inode != NULL) {
1527                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1528                         return 1;
1529         } else if (ldiskfs_has_feature_extents(sb)) {
1530                 return 1;
1531         }
1532         return 0;
1533 }
1534
1535 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1536                            const loff_t size, const loff_t pos,
1537                            const int blocks)
1538 {
1539         int credits, bits, bs, i;
1540
1541         bits = sb->s_blocksize_bits;
1542         bs = 1 << bits;
1543
1544         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1545          * we do not expect blockmaps on the large files,
1546          * so let's shrink it to 2 levels (4GB files) */
1547
1548         /* this is default reservation: 2 levels */
1549         credits = (blocks + 2) * 3;
1550
1551         /* actual offset is unknown, hard to optimize */
1552         if (pos == -1)
1553                 return credits;
1554
1555         /* now check for few specific cases to optimize */
1556         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1557                 /* no indirects */
1558                 credits = blocks;
1559                 /* allocate if not allocated */
1560                 if (inode == NULL) {
1561                         credits += blocks * 2;
1562                         return credits;
1563                 }
1564                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1565                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1566                         if (LDISKFS_I(inode)->i_data[i] == 0)
1567                                 credits += 2;
1568                 }
1569         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1570                 /* single indirect */
1571                 credits = blocks * 3;
1572                 if (inode == NULL ||
1573                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1574                         credits += 3;
1575                 else
1576                         /* The indirect block may be modified. */
1577                         credits += 1;
1578         }
1579
1580         return credits;
1581 }
1582
1583 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1584                                  const struct lu_buf *buf, loff_t _pos,
1585                                  struct thandle *handle)
1586 {
1587         struct osd_object  *obj  = osd_dt_obj(dt);
1588         struct inode       *inode = obj->oo_inode;
1589         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1590         struct osd_thandle *oh;
1591         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1592         int                 bits, bs;
1593         int                 depth, size;
1594         loff_t              pos;
1595         ENTRY;
1596
1597         LASSERT(buf != NULL);
1598         LASSERT(handle != NULL);
1599
1600         oh = container_of(handle, struct osd_thandle, ot_super);
1601         LASSERT(oh->ot_handle == NULL);
1602
1603         size = buf->lb_len;
1604         bits = sb->s_blocksize_bits;
1605         bs = 1 << bits;
1606
1607         if (_pos == -1) {
1608                 /* if this is an append, then we
1609                  * should expect cross-block record */
1610                 pos = 0;
1611         } else {
1612                 pos = _pos;
1613         }
1614
1615         /* blocks to modify */
1616         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1617         LASSERT(blocks > 0);
1618
1619         if (inode != NULL && _pos != -1) {
1620                 /* object size in blocks */
1621                 est = (i_size_read(inode) + bs - 1) >> bits;
1622                 allocated = inode->i_blocks >> (bits - 9);
1623                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1624                         /* looks like an overwrite, no need to modify tree */
1625                         credits = blocks;
1626                         /* no need to modify i_size */
1627                         goto out;
1628                 }
1629         }
1630
1631         if (osd_extents_enabled(sb, inode)) {
1632                 /*
1633                  * many concurrent threads may grow tree by the time
1634                  * our transaction starts. so, consider 2 is a min depth
1635                  * for every level we may need to allocate a new block
1636                  * and take some entries from the old one. so, 3 blocks
1637                  * to allocate (bitmap, gd, itself) + old block - 4 per
1638                  * level.
1639                  */
1640                 depth = inode != NULL ? ext_depth(inode) : 0;
1641                 depth = max(depth, 1) + 1;
1642                 credits = depth;
1643                 /* if not append, then split may need to modify
1644                  * existing blocks moving entries into the new ones */
1645                 if (_pos != -1)
1646                         credits += depth;
1647                 /* blocks to store data: bitmap,gd,itself */
1648                 credits += blocks * 3;
1649         } else {
1650                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1651         }
1652         /* if inode is created as part of the transaction,
1653          * then it's counted already by the creation method */
1654         if (inode != NULL)
1655                 credits++;
1656
1657 out:
1658
1659         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1660
1661         /* dt_declare_write() is usually called for system objects, such
1662          * as llog or last_rcvd files. We needn't enforce quota on those
1663          * objects, so always set the lqi_space as 0. */
1664         if (inode != NULL)
1665                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1666                                            i_gid_read(inode),
1667                                            i_projid_read(inode), 0,
1668                                            oh, obj, NULL, OSD_QID_BLK);
1669
1670         if (rc == 0)
1671                 rc = osd_trunc_lock(obj, oh, true);
1672
1673         RETURN(rc);
1674 }
1675
1676 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1677 {
1678         /* LU-2634: clear the extent format for fast symlink */
1679         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1680
1681         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1682         spin_lock(&inode->i_lock);
1683         LDISKFS_I(inode)->i_disksize = buflen;
1684         i_size_write(inode, buflen);
1685         spin_unlock(&inode->i_lock);
1686         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1687
1688         return 0;
1689 }
1690
1691 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1692                                     int bufsize, int write_NUL, loff_t *offs,
1693                                     handle_t *handle)
1694 {
1695         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1696         struct buffer_head *bh        = NULL;
1697         loff_t              offset    = *offs;
1698         loff_t              new_size  = i_size_read(inode);
1699         unsigned long       block;
1700         int                 blocksize = 1 << inode->i_blkbits;
1701         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1702         int                 err = 0;
1703         int                 size;
1704         int                 boffs;
1705         int                 dirty_inode = 0;
1706         bool create, sparse, sync = false;
1707
1708         if (write_NUL) {
1709                 /*
1710                  * long symlink write does not count the NUL terminator in
1711                  * bufsize, we write it, and the inode's file size does not
1712                  * count the NUL terminator as well.
1713                  */
1714                 ((char *)buf)[bufsize] = '\0';
1715                 ++bufsize;
1716         }
1717
1718         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA, &ei->i_flags);
1719
1720         /* sparse checking is racy, but sparse is very rare case, leave as is */
1721         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1722                   ((new_size - 1) >> inode->i_blkbits) + 1);
1723
1724         while (bufsize > 0) {
1725                 int credits = handle->h_buffer_credits;
1726                 unsigned long last_block = (new_size == 0) ? 0 :
1727                                            (new_size - 1) >> inode->i_blkbits;
1728
1729                 if (bh)
1730                         brelse(bh);
1731
1732                 block = offset >> inode->i_blkbits;
1733                 boffs = offset & (blocksize - 1);
1734                 size = min(blocksize - boffs, bufsize);
1735                 sync = (block > last_block || new_size == 0 || sparse);
1736
1737                 if (sync)
1738                         down(&ei->i_append_sem);
1739
1740                 bh = __ldiskfs_bread(handle, inode, block, 0);
1741
1742                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1743                         CWARN("%s: adding bh without locking off %llu (block %lu, "
1744                               "size %d, offs %llu)\n", inode->i_sb->s_id,
1745                               offset, block, bufsize, *offs);
1746
1747                 if (IS_ERR_OR_NULL(bh)) {
1748                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1749                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1750
1751                         /* while the file system is being mounted, avoid
1752                          * preallocation otherwise mount can take a long
1753                          * time as mballoc cache is cold.
1754                          * XXX: this is a workaround until we have a proper
1755                          *      fix in mballoc
1756                          * XXX: works with extent-based files only */
1757                         if (!osd->od_cl_seq)
1758                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1759                         bh = __ldiskfs_bread(handle, inode, block, flags);
1760                         create = true;
1761                 } else {
1762                         if (sync) {
1763                                 up(&ei->i_append_sem);
1764                                 sync = false;
1765                         }
1766                         create = false;
1767                 }
1768                 if (IS_ERR_OR_NULL(bh)) {
1769                         if (bh == NULL) {
1770                                 err = -EIO;
1771                         } else {
1772                                 err = PTR_ERR(bh);
1773                                 bh = NULL;
1774                         }
1775
1776                         CERROR("%s: error reading offset %llu (block %lu, "
1777                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1778                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1779                                credits, handle->h_buffer_credits, err);
1780                         break;
1781                 }
1782
1783                 err = ldiskfs_journal_get_write_access(handle, bh);
1784                 if (err) {
1785                         CERROR("journal_get_write_access() returned error %d\n",
1786                                err);
1787                         break;
1788                 }
1789                 LASSERTF(boffs + size <= bh->b_size,
1790                          "boffs %d size %d bh->b_size %lu\n",
1791                          boffs, size, (unsigned long)bh->b_size);
1792                 if (create) {
1793                         memset(bh->b_data, 0, bh->b_size);
1794                         if (sync) {
1795                                 up(&ei->i_append_sem);
1796                                 sync = false;
1797                         }
1798                 }
1799                 memcpy(bh->b_data + boffs, buf, size);
1800                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1801                 if (err)
1802                         break;
1803
1804                 if (offset + size > new_size)
1805                         new_size = offset + size;
1806                 offset += size;
1807                 bufsize -= size;
1808                 buf += size;
1809         }
1810         if (sync)
1811                 up(&ei->i_append_sem);
1812
1813         if (bh)
1814                 brelse(bh);
1815
1816         if (write_NUL)
1817                 --new_size;
1818         /* correct in-core and on-disk sizes */
1819         if (new_size > i_size_read(inode)) {
1820                 spin_lock(&inode->i_lock);
1821                 if (new_size > i_size_read(inode))
1822                         i_size_write(inode, new_size);
1823                 if (i_size_read(inode) > ei->i_disksize) {
1824                         ei->i_disksize = i_size_read(inode);
1825                         dirty_inode = 1;
1826                 }
1827                 spin_unlock(&inode->i_lock);
1828         }
1829         if (dirty_inode)
1830                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1831
1832         if (err == 0)
1833                 *offs = offset;
1834         return err;
1835 }
1836
1837 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1838                          const struct lu_buf *buf, loff_t *pos,
1839                          struct thandle *handle)
1840 {
1841         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1842         struct osd_thandle      *oh;
1843         ssize_t                 result;
1844         int                     is_link;
1845
1846         LASSERT(dt_object_exists(dt));
1847
1848         LASSERT(handle != NULL);
1849         LASSERT(inode != NULL);
1850         dquot_initialize(inode);
1851
1852         /* XXX: don't check: one declared chunk can be used many times */
1853         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1854
1855         oh = container_of(handle, struct osd_thandle, ot_super);
1856         LASSERT(oh->ot_handle->h_transaction != NULL);
1857         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1858
1859         /* Write small symlink to inode body as we need to maintain correct
1860          * on-disk symlinks for ldiskfs.
1861          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1862          * does not count it in.
1863          */
1864         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1865         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1866                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1867         else
1868                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1869                                                   is_link, pos, oh->ot_handle);
1870         if (result == 0)
1871                 result = buf->lb_len;
1872
1873         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1874
1875         return result;
1876 }
1877
1878 static int osd_declare_fallocate(const struct lu_env *env,
1879                                  struct dt_object *dt, struct thandle *th)
1880 {
1881         struct osd_thandle *oh;
1882         struct inode *inode;
1883         int rc;
1884         ENTRY;
1885
1886         LASSERT(th);
1887         oh = container_of(th, struct osd_thandle, ot_super);
1888
1889         osd_trans_declare_op(env, oh, OSD_OT_PREALLOC,
1890                              osd_dto_credits_noquota[DTO_WRITE_BLOCK]);
1891         inode = osd_dt_obj(dt)->oo_inode;
1892         LASSERT(inode);
1893
1894         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1895                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1896                                    NULL, OSD_QID_BLK);
1897         RETURN(rc);
1898 }
1899
1900 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1901                          __u64 start, __u64 end, int mode, struct thandle *th)
1902 {
1903         struct osd_object *obj = osd_dt_obj(dt);
1904         struct inode *inode = obj->oo_inode;
1905         int rc = 0;
1906         struct osd_thread_info *info = osd_oti_get(env);
1907         struct dentry *dentry = &info->oti_obj_dentry;
1908         struct file *file = &info->oti_file;
1909
1910         ENTRY;
1911         /*
1912          * Only mode == 0 (which is standard prealloc) is supported now.
1913          * Rest of mode options is not supported yet.
1914          */
1915         if (mode & ~FALLOC_FL_KEEP_SIZE)
1916                 RETURN(-EOPNOTSUPP);
1917
1918         LASSERT(dt_object_exists(dt));
1919         LASSERT(osd_invariant(obj));
1920         LASSERT(inode != NULL);
1921         dquot_initialize(inode);
1922
1923         LASSERT(th);
1924
1925         osd_trans_exec_op(env, th, OSD_OT_PREALLOC);
1926
1927         /*
1928          * Because f_op->fallocate() does not have an inode arg
1929          */
1930         dentry->d_inode = inode;
1931         dentry->d_sb = inode->i_sb;
1932         file->f_path.dentry = dentry;
1933         file->f_mapping = inode->i_mapping;
1934         file->f_op = inode->i_fop;
1935         file->f_inode = inode;
1936         rc = file->f_op->fallocate(file, mode, start, end - start);
1937
1938         RETURN(rc);
1939 }
1940
1941 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1942                              __u64 start, __u64 end, struct thandle *th)
1943 {
1944         struct osd_thandle *oh;
1945         struct inode       *inode;
1946         int                 rc;
1947         ENTRY;
1948
1949         LASSERT(th);
1950         oh = container_of(th, struct osd_thandle, ot_super);
1951
1952         /*
1953          * we don't need to reserve credits for whole truncate
1954          * it's not possible as truncate may need to free too many
1955          * blocks and that won't fit a single transaction. instead
1956          * we reserve credits to change i_size and put inode onto
1957          * orphan list. if needed truncate will extend or restart
1958          * transaction
1959          */
1960         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1961                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1962
1963         inode = osd_dt_obj(dt)->oo_inode;
1964         LASSERT(inode);
1965
1966         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1967                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1968                                    NULL, OSD_QID_BLK);
1969
1970         if (rc == 0)
1971                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1972
1973         RETURN(rc);
1974 }
1975
1976 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1977                      __u64 start, __u64 end, struct thandle *th)
1978 {
1979         struct osd_object *obj = osd_dt_obj(dt);
1980         struct osd_device *osd = osd_obj2dev(obj);
1981         struct inode *inode = obj->oo_inode;
1982         struct osd_access_lock *al;
1983         struct osd_thandle *oh;
1984         int rc = 0, found = 0;
1985         bool grow = false;
1986         ENTRY;
1987
1988         LASSERT(dt_object_exists(dt));
1989         LASSERT(osd_invariant(obj));
1990         LASSERT(inode != NULL);
1991         dquot_initialize(inode);
1992
1993         LASSERT(th);
1994         oh = container_of(th, struct osd_thandle, ot_super);
1995         LASSERT(oh->ot_handle->h_transaction != NULL);
1996
1997         /* we used to skip truncate to current size to
1998          * optimize truncates on OST. with DoM we can
1999          * get attr_set to set specific size (MDS_REINT)
2000          * and then get truncate RPC which essentially
2001          * would be skipped. this is bad.. so, disable
2002          * this optimization on MDS till the client stop
2003          * to sent MDS_REINT (LU-11033) -bzzz */
2004         if (osd->od_is_ost && i_size_read(inode) == start)
2005                 RETURN(0);
2006
2007         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2008
2009         spin_lock(&inode->i_lock);
2010         if (i_size_read(inode) < start)
2011                 grow = true;
2012         i_size_write(inode, start);
2013         spin_unlock(&inode->i_lock);
2014         /* if object holds encrypted content, we need to make sure we truncate
2015          * on an encryption unit boundary, or subsequent reads will get
2016          * corrupted content
2017          */
2018         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2019             start & ~LUSTRE_ENCRYPTION_MASK)
2020                 start = (start & LUSTRE_ENCRYPTION_MASK) +
2021                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2022         ll_truncate_pagecache(inode, start);
2023
2024         /* optimize grow case */
2025         if (grow) {
2026                 osd_execute_truncate(obj);
2027                 GOTO(out, rc);
2028         }
2029
2030         inode_lock(inode);
2031         /* add to orphan list to ensure truncate completion
2032          * if this transaction succeed. ldiskfs_truncate()
2033          * will take the inode out of the list */
2034         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2035         inode_unlock(inode);
2036         if (rc != 0)
2037                 GOTO(out, rc);
2038
2039         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2040                 if (obj != al->tl_obj)
2041                         continue;
2042                 LASSERT(al->tl_shared == 0);
2043                 found = 1;
2044                 /* do actual truncate in osd_trans_stop() */
2045                 al->tl_truncate = 1;
2046                 break;
2047         }
2048         LASSERT(found);
2049
2050 out:
2051         RETURN(rc);
2052 }
2053
2054 static int fiemap_check_ranges(struct inode *inode,
2055                                u64 start, u64 len, u64 *new_len)
2056 {
2057         loff_t maxbytes;
2058
2059         *new_len = len;
2060
2061         if (len == 0)
2062                 return -EINVAL;
2063
2064         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2065                 maxbytes = inode->i_sb->s_maxbytes;
2066         else
2067                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2068
2069         if (start > maxbytes)
2070                 return -EFBIG;
2071
2072         /*
2073          * Shrink request scope to what the fs can actually handle.
2074          */
2075         if (len > maxbytes || (maxbytes - len) < start)
2076                 *new_len = maxbytes - start;
2077
2078         return 0;
2079 }
2080
2081 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2082 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2083
2084 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2085                           struct fiemap *fm)
2086 {
2087         struct fiemap_extent_info fieinfo = {0, };
2088         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2089         u64 len;
2090         int rc;
2091         mm_segment_t cur_fs;
2092
2093         LASSERT(inode);
2094         if (inode->i_op->fiemap == NULL)
2095                 return -EOPNOTSUPP;
2096
2097         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2098                 return -EINVAL;
2099
2100         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2101         if (rc)
2102                 return rc;
2103
2104         fieinfo.fi_flags = fm->fm_flags;
2105         fieinfo.fi_extents_max = fm->fm_extent_count;
2106         fieinfo.fi_extents_start = fm->fm_extents;
2107
2108         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2109                 filemap_write_and_wait(inode->i_mapping);
2110
2111         /* Save previous value address limit */
2112         cur_fs = get_fs();
2113         /* Set the address limit of the kernel */
2114         set_fs(KERNEL_DS);
2115
2116         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2117         fm->fm_flags = fieinfo.fi_flags;
2118         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2119
2120         /* Restore the previous address limt */
2121         set_fs(cur_fs);
2122
2123         return rc;
2124 }
2125
2126 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2127                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2128 {
2129         struct osd_object *obj = osd_dt_obj(dt);
2130         int rc = 0;
2131         ENTRY;
2132
2133         switch (advice) {
2134         case LU_LADVISE_DONTNEED:
2135                 if (end)
2136                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2137                                                  start >> PAGE_SHIFT,
2138                                                  (end - 1) >> PAGE_SHIFT);
2139                 break;
2140         default:
2141                 rc = -ENOTSUPP;
2142                 break;
2143         }
2144
2145         RETURN(rc);
2146 }
2147
2148 /*
2149  * in some cases we may need declare methods for objects being created
2150  * e.g., when we create symlink
2151  */
2152 const struct dt_body_operations osd_body_ops_new = {
2153         .dbo_declare_write = osd_declare_write,
2154 };
2155
2156 const struct dt_body_operations osd_body_ops = {
2157         .dbo_read                       = osd_read,
2158         .dbo_declare_write              = osd_declare_write,
2159         .dbo_write                      = osd_write,
2160         .dbo_bufs_get                   = osd_bufs_get,
2161         .dbo_bufs_put                   = osd_bufs_put,
2162         .dbo_write_prep                 = osd_write_prep,
2163         .dbo_declare_write_commit       = osd_declare_write_commit,
2164         .dbo_write_commit               = osd_write_commit,
2165         .dbo_read_prep                  = osd_read_prep,
2166         .dbo_declare_punch              = osd_declare_punch,
2167         .dbo_punch                      = osd_punch,
2168         .dbo_fiemap_get                 = osd_fiemap_get,
2169         .dbo_ladvise                    = osd_ladvise,
2170         .dbo_declare_fallocate          = osd_declare_fallocate,
2171         .dbo_fallocate                  = osd_fallocate,
2172 };
2173
2174 /**
2175  * Get a truncate lock
2176  *
2177  * In order to take multi-transaction truncate out of main transaction we let
2178  * the caller grab a lock on the object passed. the lock can be shared (for
2179  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2180  * and write in the same transaction handle (do not confuse with big ldiskfs
2181  * transaction containing lots of handles).
2182  * The lock must be taken at declaration.
2183  *
2184  * \param obj           object to lock
2185  * \oh                  transaction
2186  * \shared              shared or exclusive
2187  *
2188  * \retval 0            lock is granted
2189  * \retval -NOMEM       no memory to allocate lock
2190  */
2191 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2192 {
2193         struct osd_access_lock *al, *tmp;
2194
2195         LASSERT(obj);
2196         LASSERT(oh);
2197
2198         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2199                 if (tmp->tl_obj != obj)
2200                         continue;
2201                 LASSERT(tmp->tl_shared == shared);
2202                 /* found same lock */
2203                 return 0;
2204         }
2205
2206         OBD_ALLOC_PTR(al);
2207         if (unlikely(al == NULL))
2208                 return -ENOMEM;
2209         al->tl_obj = obj;
2210         al->tl_truncate = false;
2211         if (shared)
2212                 down_read(&obj->oo_ext_idx_sem);
2213         else
2214                 down_write(&obj->oo_ext_idx_sem);
2215         al->tl_shared = shared;
2216         lu_object_get(&obj->oo_dt.do_lu);
2217
2218         list_add(&al->tl_list, &oh->ot_trunc_locks);
2219
2220         return 0;
2221 }
2222
2223 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2224 {
2225         struct osd_access_lock *al, *tmp;
2226         list_for_each_entry_safe(al, tmp, list, tl_list) {
2227                 if (al->tl_shared)
2228                         up_read(&al->tl_obj->oo_ext_idx_sem);
2229                 else
2230                         up_write(&al->tl_obj->oo_ext_idx_sem);
2231                 osd_object_put(env, al->tl_obj);
2232                 list_del(&al->tl_list);
2233                 OBD_FREE_PTR(al);
2234         }
2235 }
2236
2237 void osd_execute_truncate(struct osd_object *obj)
2238 {
2239         struct osd_device *d = osd_obj2dev(obj);
2240         struct inode *inode = obj->oo_inode;
2241         __u64 size;
2242
2243         /* simulate crash before (in the middle) of delayed truncate */
2244         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2245                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2246                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2247
2248                 mutex_lock(&sbi->s_orphan_lock);
2249                 list_del_init(&ei->i_orphan);
2250                 mutex_unlock(&sbi->s_orphan_lock);
2251                 return;
2252         }
2253
2254         size = i_size_read(inode);
2255         inode_lock(inode);
2256         /* if object holds encrypted content, we need to make sure we truncate
2257          * on an encryption unit boundary, or block content will get corrupted
2258          */
2259         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2260             size & ~LUSTRE_ENCRYPTION_MASK)
2261                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2262                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2263         ldiskfs_truncate(inode);
2264         inode_unlock(inode);
2265         if (inode->i_size != size) {
2266                 spin_lock(&inode->i_lock);
2267                 i_size_write(inode, size);
2268                 LDISKFS_I(inode)->i_disksize = size;
2269                 spin_unlock(&inode->i_lock);
2270                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2271         }
2272
2273         /*
2274          * For a partial-page truncate, flush the page to disk immediately to
2275          * avoid data corruption during direct disk write.  b=17397
2276          */
2277         if ((size & ~PAGE_MASK) == 0)
2278                 return;
2279         if (osd_use_page_cache(d)) {
2280                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2281         } else {
2282                 /* Notice we use "wait" version to ensure I/O is complete */
2283                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2284                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2285                                          size >> PAGE_SHIFT);
2286         }
2287 }
2288
2289 void osd_process_truncates(struct list_head *list)
2290 {
2291         struct osd_access_lock *al;
2292
2293         LASSERT(journal_current_handle() == NULL);
2294
2295         list_for_each_entry(al, list, tl_list) {
2296                 if (al->tl_shared)
2297                         continue;
2298                 if (!al->tl_truncate)
2299                         continue;
2300                 osd_execute_truncate(al->tl_obj);
2301         }
2302 }